Scientists discover surprising language 'shortcuts' in birdsong – like humans
36 points
4 days ago
| 2 comments
| manchester.ac.uk
| HN
orwin
9 hours ago
[-]
Four years ago, I wrote that to me, Transformers' most exiting application could be translating whale's songs. I was obviously very wrong (won't be the first time, won't be the last), but I imagine recording of birds should be more numerous than whales', so maybe someday, hopefully, transformers would integrate bird songs to the list of languages they can translate to and from.
reply
moi2388
8 hours ago
[-]
How will you train this? I understand it replicating bird songs, but what data will it use for actual translation?
reply
tomrod
5 hours ago
[-]
It would operate a lot like image recognition tasks, probably in frequency 0 domain for a concrete space to operate in. SLID or other information theoretic approaches could isolate common signals, then translate that to captured environmental information (e.g prairie dogs identifying predators).

Animals don't use a known syntax per se, so it wouldn't be authentic translation, but a transliteration may be possible. Also, there is no guarantee that one animal doing something (like a dog's behavior for going to the bathroom) maps to many or all dogs.

reply
chrisco255
8 hours ago
[-]
The Crowsetta Stone
reply
crowsettastone
5 hours ago
[-]
reply
imglorp
4 hours ago
[-]
Amazing, I gave it a recording and it gave an analysis.

Do you know how to hear the generated sounds? Pressing the buttons shows a "playing" console message but there's no audio.

reply
bhickey
4 hours ago
[-]
Probably the same way many other models are trained: autoregression or autoencoding. You'd either predict the next symbol, or you compress it into a latent space and reconstruct the original. My assumption is that birdsong is sequential, but this isn't something I know about. An entire song might be the smallest semantic unit, like the alien language in Arrival, though I think this is unlike.
reply
diggan
6 hours ago
[-]
> but what data will it use for actual translation?

I'm no ornithologist, for forgive my ignorance, but shouldn't there be bunch of papers out there where researchers try to infer the meaning of various sounds birds produce, together with a description of the sound and even samples of it? I don't know how numerous that could be, but could maybe be used as a starting point at least.

reply
ileonichwiesz
5 hours ago
[-]
There are large datasets of bird sounds (eBird, Cornell Labs, etc), but the descriptions are usually limited to the species, location, and something like “mating call” or “contact call”. Hugely useful for building models that can recognise birds by call (apps like that already exist, I recommend Merlin Bird ID), but definitely not enough for something approaching actual translation.

FWIW there’s no research to indicate that the sounds birds make are what we’d call language. They’re avid communicators, and some species are known to be highly intelligent, but of course it’s not like “caw” means “to fly” and “craah” is “forward”.

reply
macrolime
7 hours ago
[-]
YouTube videos of birds?
reply
loveparade
7 hours ago
[-]
I hope the EditorBird added subtitles.
reply
pjmlp
6 hours ago
[-]
I would actually be quite curious what the crows get to talk around 6 AM on nearby trees for about half an hour, before leaving for their daily duties.

Not sure how everything would go, if we finally managed to talk with fellow species, given how wrong it goes even with other humans.

reply
jacquesm
5 hours ago
[-]
That's the country wide weather report. And they need to recite it so whole that it spreads from the point of origin otherwise it would be just a local update. Fascinating stuff. Humans used to do much the same when radio first came along.
reply
pjmlp
5 hours ago
[-]
Lovely. :)
reply
jacquesm
4 hours ago
[-]
Plot twist: 50 years from now they really decode it and it turns out to be exactly that. I'll be eating some crow if that should be the case.
reply
suddenlybananas
8 hours ago
[-]
Don't monkeys on a typewriter with a space bar also have the same statistical property?
reply
namenotrequired
8 hours ago
[-]
That’s why the researchers used

> … a new open-source computational tool called ZLAvian, which compares real-world observed patterns to simulated ones to determine if ZLA is present.

reply
suddenlybananas
7 hours ago
[-]
I'm not doubting that the finding is there, I'm expressing scepticism that it means very much. If you randomly sample uniformly from the set {"a", "b", " "} repeatedly, the "word "ab" will appear much more often than the "word", "aaaaaabbbbababa". Doesn't say very much about language itself though.
reply
mannykannot
7 hours ago
[-]
You got me interested in the question, and one of the first things I found out is that there is Zipf’s Law, and then there is Zipf’s Law of Abbreviation.

Monkeys-randomly-typing is one of many processes which will indeed generate sequences conforming to the former, and it is perhaps the exceptions which are most interesting.

The latter law observes that the former generally applies to sequences generated for communication, having semantics and usually a grammar. While this may be the expected finding, there is value in having this expectation empirically verified.

reply
INTPenis
6 hours ago
[-]
Interesting comment since this is from the university of Manchester. Is this a late return to monkey news?
reply